People say AI agents don't work yet. They point to hallucinations, runaway costs, the occasional disaster demo. All real. But they're missing the deeper problem.

The word itself betrays us. We keep treating agents like advanced tools. A hammer doesn't need a to-do list. A search engine doesn't decide to work on something else halfway through your query. You give a tool an input, you get an output. Simple.

Agents aren't tools. They're decision-making systems wrapped in language. And every decision they make — routing a payment, joining a group chat, prioritizing one task over another — multiplies the complexity exponentially. The model isn't the bottleneck. Managing the fact that the thing has opinions about what to do next is.

The people building useful agents stop trying to make the models smarter. They focus on constraints, handoff protocols, and the stuff nobody writes papers about: making sure your agent knows when to shut up.

When did we start building systems that need judgment and then act surprised when they need judgment to function?